Rhoda AI, a robotics research company, says the size and training budget of the video model behind a robot’s control system predicts how well that robot performs a real industrial job better than simply collecting more demonstrations does. The company published the finding in a research post describing its own internal study, so the result is a self-reported claim rather than an independently verified one. Still, the stakes are real: robotics teams currently spend heavily on gathering robot-specific demonstration data, and Rhoda’s study argues that money might be better aimed at the video model itself.
The systems Rhoda tested are what the company calls Direct Video-Action models, or DVA. A DVA starts as a causal video model that learns to predict how a scene will unfold after training on ordinary web video, none of it showing robots or manipulation tasks and none of it labeled with actions. Only afterward does the company “post-train” that model on demonstrations of one specific job. A separate inverse dynamics model then converts the video model’s predicted future frames into the motor commands a robot arm actually executes.
The wager behind this approach is that a system exposed to enormous volumes of ordinary footage, hands cooking, people assembling furniture, objects being picked up and set down, absorbs some general sense of how physical objects behave. That sense, the theory goes, should transfer to a task the model has never specifically seen, cutting the amount of expensive, operator-collected robot data a team needs before a policy works.
Rhoda tested the idea on a real customer task: unpacking bearings from boxes and sorting the packaging waste, a long, multi-step job that takes over a minute even for a skilled human operator. The company ran more than 200 hours of real-robot evaluation, hundreds of trials per policy, scoring each one on “at-speed completion rate,” the share of trials finished correctly and within a strict time limit.
Three findings anchor the post. First, larger pretrained video models produced better robot policies at every size tested, from an extra-small model that completed just 4% of trials up to a large model at 85%, with the improvement still climbing at the largest size Rhoda trained. Second, more pretraining compute helped at every amount of post-training robot data the company tried, and the effect was largest when that robot data was scarce: three compute budgets of the same mid-size model landed within 8 points of each other with a full demonstration set, but 31 points apart when post-trained on just a quarter of it. Third, a metric Rhoda calls DINO FD (Frechet distance) tracked task performance across both experiments. It works by having a pretrained video model predict how held-out web video clips continue, then measuring the statistical distance between its predicted frames and the real ones, using a frozen DINOv2 image encoder to turn frames into comparable features. Lower distance means the model’s predictions match reality more closely, and Rhoda found that models with a lower DINO FD score consistently produced better robot policies, regardless of whether that score came from a bigger model or a longer training run.
Rhoda’s own paper flags real limits: the results come from one task, one robot, and a single training run per condition, so the correlation between pretraining quality and robot performance is not yet established as causal. No outside lab has replicated the setup, and the company has an obvious interest in a result that favors its own architecture.
For teams building or buying robot foundation models, the practical read is a capital allocation question familiar from large language models: pretraining compute is a cost paid once and reused across every future task, while demonstration data is a cost paid again for each new site and product. Rhoda’s data suggests that trade favors pretraining scale whenever a team’s task-specific data is limited, and operators evaluating robotics vendors should ask whether a system’s advantage comes from its video pretraining or simply from more collected demonstrations.
Findings reported by Rhoda AI in a research post on its own website, published 2026.
